mersenneforum.org  

Go Back   mersenneforum.org > Great Internet Mersenne Prime Search > Data

Reply
 
Thread Tools
Old 2019-07-02, 05:11   #199
SethTro
 
SethTro's Avatar
 
"Seth"
Apr 2019

2·103 Posts
Default

Quote:
Originally Posted by SethTro View Post
I read about 80% of this thread and tried to follow all the math based suggestions.

@Preda's posts, especially 58 and 122, are great.
https://www.mersenneforum.org/showpo...7&postcount=58
https://www.mersenneforum.org/showpo...&postcount=122

I had the same ideas after reading the initial idea by Robert Gerbicz
https://www.mersenneforum.org/showpo...7&postcount=30

I have another adapting that I believe can solve one of the listed drawbacks



We can also adapt a solution used by bitcoin mining pools (where they similarly miners to have success mining (e.g. "see a factor") only rarely).

Instead of giving binary all or nothing credit (and hoping it averages over a long period of timing) give credit proportional to the smallness of min(x) or min(bitrev(x). Now each result earns between 1/32 and ~4 credit on each submission. This function is much less noisy than the binary 0/1 reward so even when turning in a bunch of no factor found results you would still expect to receive, on average, full credit (I can write a simulator if people want graphs).

Now hypothetical cheater would get credit exactly equal to the work done on average. It doesn't stop them lying about the work done (that can maybe be detected later with static user analysis) but they're reward is now verifiable linked to the work they performed
I wrote up a 80% of the code needed for this
https://github.com/sethtroisi/mfaktc/tree/master

mod_simple_96_and_check_big_factor96 doesn't calculate the modulo so it uses a slightly different Proof-of-work function (instead producing very large modulos instead of small)
https://github.com/sethtroisi/mfaktc...helper.cu#L244

The end result is on TF-NF results you get a little extra line like this
Code:
M59068201 proof_k(17257705361971287559): 30 bits [TF:60:64:mfaktc 0.21 75bit_mul32_gs]
M59068201 proof_k(1759939290551364353): 31 bits [TF:60:64:mfaktc 0.21 75bit_mul32_gs]
M59068201 proof_k(1297657372566442343): 31 bits [TF:60:64:mfaktc 0.21 75bit_mul32_gs]
M59068201 proof_k(8940824503190951977): 29 bits [TF:60:64:mfaktc 0.21 75bit_mul32_gs]
[Mon Jul  1 21:59:54 2019]
no factor for M59068201 from 2^60 to 2^64 [mfaktc 0.21 75bit_mul32_gs]
You can then verify that pow(2, 59068201, 8940824503190951977) = 422536362 which is 29 bits meaning ~34 leading zeros bits. so it takes around ~1e11 test to find this (which is handily just about 2^64 / 59068201)
SethTro is offline   Reply With Quote
Old 2020-01-09, 01:29   #200
SethTro
 
SethTro's Avatar
 
"Seth"
Apr 2019

2·103 Posts
Default

What's the appropriate way to report probable false results?

M14951 has a P-1 result with B2=9,887,122,214,540,712 which took an estimated 186,161 GHzDays (I guess this isn't impossible, but seems unlikely given the user doesn't appear on any of the top producer lists are regularly submit factors)
SethTro is offline   Reply With Quote
Old 2020-01-09, 02:48   #201
axn
 
axn's Avatar
 
Jun 2003

4,789 Posts
Default

Quote:
Originally Posted by SethTro View Post
What's the appropriate way to report probable false results?

M14951 has a P-1 result with B2=9,887,122,214,540,712 which took an estimated 186,161 GHzDays (I guess this isn't impossible, but seems unlikely given the user doesn't appear on any of the top producer lists are regularly submit factors)
This is a perfectly doable range if you use P95 for stage1 and GMP-ECM for stage 2. The 186k GHzdays is granted based on the assumption that stage 2 was done with P95.
axn is offline   Reply With Quote
Old 2020-01-09, 04:24   #202
SethTro
 
SethTro's Avatar
 
"Seth"
Apr 2019

2×103 Posts
Default

Quote:
Originally Posted by axn View Post
This is a perfectly doable range if you use P95 for stage1 and GMP-ECM for stage 2. The 186k GHzdays is granted based on the assumption that stage 2 was done with P95.
It sticks out quite a bit on the factoring limit page. with B2=9e15 it's the 2nd largest B2 and 2-3 orders of magnitude larger than the B2 of nearby exponents (M10061 had B2=1e14, M16411 had B2=5e12)

https://www.mersenne.org/report_fact...99&tftobits=72
SethTro is offline   Reply With Quote
Old 2020-01-09, 06:38   #203
axn
 
axn's Avatar
 
Jun 2003

4,789 Posts
Default

Quote:
Originally Posted by SethTro View Post
It sticks out quite a bit on the factoring limit page. with B2=9e15 it's the 2nd largest B2 and 2-3 orders of magnitude larger than the B2 of nearby exponents (M10061 had B2=1e14, M16411 had B2=5e12)

https://www.mersenne.org/report_fact...99&tftobits=72
From the ratio of B2/B1, it looks like most of them were done with P95 for both stages. Couple of them, 2719 and 14951 sticks out.

Like I said, this is perfectly normal _if_ GMP ECM was used for stage 2. Unless you have some reason to believe that these are specifically fraudulent (apart from the large B2), I suggest that you make peace with it.
axn is offline   Reply With Quote
Old 2020-04-28, 16:12   #204
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

32·232 Posts
Default Verifying TF was done

There's a new discussion of verification in https://mersenneforum.org/showthread.php?t=25493
kriesel is offline   Reply With Quote
Old 2020-08-08, 00:31   #205
lycorn
 
lycorn's Avatar
 
Sep 2002
Oeiras, Portugal

24×89 Posts
Default

I reckon we have another potential case of TF cheating / defective hardware.
User Dirk (#21 in top TF list) has 727,087 GHzD credited for 228 trials and 0 factors found. This amounts to an average of 3,189 GHzD per trial.
Today I managed to track some of the recent trials, and the last 4 were in the low 200M range, from 2^71 to 2^85 (!). I don´t know about the previous ones, but I find these results rather suspicious. Would someone with privileged access to the server bother to have a look at it?
lycorn is offline   Reply With Quote
Old 2020-08-08, 01:36   #206
James Heinrich
 
James Heinrich's Avatar
 
"James Heinrich"
May 2004
ex-Northern Ontario

3·11·97 Posts
Default

Quote:
Originally Posted by lycorn View Post
I reckon we have another potential case of TF cheating / defective hardware. User Dirk (#21 in top TF list) has 727,087 GHzD credited for 228 trials and 0 factors found. This amounts to an average of 3,189 GHzD per trial. Today I managed to track some of the recent trials, and the last 4 were in the low 200M range, from 2^71 to 2^85 (!). I don´t know about the previous ones, but I find these results rather suspicious. Would someone with privileged access to the server bother to have a look at it?
I took a look at his 145 results from 2020:
Code:
2020-01-29 M194726743 72-85 = 40,234 GHd
2020-02-04 M184837931 81-85 = 39,742 GHd
2020-02-26 M194726827 72-85 = 40,234 GHd
2020-03-28 M196745771 73-85 = 39,816 GHd
2020-05-13 M200434841 73-85 = 39,084 GHd
2020-05-17 M201360557 73-85 = 38,904 GHd
2020-06-30 M202932731 72-85 = 38,607 GHd
2020-06-30 M203782673 72-85 = 38,446 GHd
2020-08-07 M200687659 71-85 = 39,042 GHd
2020-08-07 M204629281 71-85 = 38,285 GHd
2020-08-07 M204629309 71-85 = 38,285 GHd
2020-08-07 M205132043 71-85 = 38,191 GHd
It's an unusual work distribution to be sure, but it could be legitimate. It works out to a bit over 2000 GHz-days of work per day, certainly achievable with a single RTX 2080. Cumulative probability for each 7?-85 run on an exponent gives a 16-19% probability of finding a factor. Not finding one over a dozen specific exponents could just be bad luck.

If it were my call I would suggest that TF credit be capped at the target TF level for that exponent (78-79 for exponents in this range), higher TF results could be accepted but no credit given since (as far as the GIMPS project is concerned) it's a waste of resources.

I'll pass this along to George to see if he wants to investigate further, perhaps talk to "Dirk" about it.
James Heinrich is offline   Reply With Quote
Old 2020-08-08, 02:57   #207
Madpoo
Serpentine Vermin Jar
 
Madpoo's Avatar
 
Jul 2014

29×113 Posts
Default

Quote:
Originally Posted by lycorn View Post
I reckon we have another potential case of TF cheating / defective hardware.
User Dirk (#21 in top TF list) has 727,087 GHzD credited for 228 trials and 0 factors found. This amounts to an average of 3,189 GHzD per trial.
Today I managed to track some of the recent trials, and the last 4 were in the low 200M range, from 2^71 to 2^85 (!). I don´t know about the previous ones, but I find these results rather suspicious. Would someone with privileged access to the server bother to have a look at it?
A cursory look shows that it's probably fine. For example:
M160173253

That's the only "factor found" result, but it is in that high bit range.

The user also has a number of PRP (M78949933) and LL tests (M99688541), as well as a few P-1 tests, sometimes on the same exponent:
M107772919

For whatever reason, this user likes to factor to high TF bit levels and that's okay.
Madpoo is offline   Reply With Quote
Old 2020-08-08, 05:52   #208
Prime95
P90 years forever!
 
Prime95's Avatar
 
Aug 2002
Yeehaw, FL

52·172 Posts
Default

Lifetime of 6 factors found:

Code:
M160080763 has a factor: 6027040288606836258928409 [TF:82:83:mfaktc 0.21 barrett87_mul32_gs]
M160173253 has a factor: 11077083125761109161661041 [TF:83:84:mfaktc 0.21 barrett87_mul32_gs]
M158139979 has a factor: 1342523932462359763583 [TF:70:71:mfaktc 0.21 barrett76_mul32_gs]
M157400489 has a factor: 3482434867307359525409 [TF:71:72:mfaktc 0.21 barrett76_mul32_gs]
M157400641 has a factor: 32396470116696604575641 [TF:74:75:mfaktc 0.21 barrett76_mul32_gs]
M160080541 has a factor: 2709190282885551671064721 [TF:81:82:mfaktc 0.21 barrett87_mul32_gs]
Prime95 is offline   Reply With Quote
Old 2020-08-08, 14:08   #209
lycorn
 
lycorn's Avatar
 
Sep 2002
Oeiras, Portugal

24×89 Posts
Default

Fair enough. Thanks for your prompt replies.
lycorn is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
Results ET_ Operazione Doppi Mersennes 622 2020-11-21 09:36
Where have all the TF results gone?... lycorn PrimeNet 22 2017-10-02 02:40
PGS Results danaj Prime Gap Searches 0 2017-08-14 18:35
CPU Results last 24 hrs Unregistered Information & Answers 3 2010-07-26 00:49
0x results... Mike PrimeNet 11 2004-05-23 12:55

All times are UTC. The time now is 02:09.

Fri Dec 4 02:09:43 UTC 2020 up 22:21, 1 user, load averages: 1.41, 1.63, 1.76

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2020, Jelsoft Enterprises Ltd.

This forum has received and complied with 0 (zero) government requests for information.

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.
A copy of the license is included in the FAQ.